Dissertation and Thesis

Now showing 1 - 5 of 5

Comparative Analysis on Different Feature Selection
(Indian Statistical Institute, Kolkata, 2024-07) Goswami, Santanu
In this research, we propose a comprehensive framework for uncovering hidden patterns, selecting optimal features, and reducing dimensionality in large datasets, particularly focusing on 10K x 10K dimensional data. Traditional methods often struggle to efficiently handle such vast datasets due to computational constraints and information overload. To address this challenge, we introduce three innovative approaches leveraging deep neural networks (DNNs) and recurrent neural networks (RNNs) to enhance pattern identification, feature selection, and dimensionality reduction. Firstly, we develop a DNN-based framework tailored to identifying hidden patterns within extensive datasets. By harnessing the representational power of deep neural networks, our framework systematically uncovers intricate relationships and structures among observations, allowing for the extraction and preservation of unique patterns for future use. Secondly, we propose an optimal feature selection framework designed to efficiently navigate through the entire feature set and identify the most informative subset. Leveraging advanced optimization techniques, our approach intelligently selects features that maximize predictive performance while minimizing redundancy, thus enhancing model interpretability and computational efficiency. Thirdly, we introduce an autoencoder-based dimension reduction method aimed at effectively reducing the dimensionality of the dataset without sacrificing crucial information. By employing the encoding phase of an autoencoder architecture, we compress the input data into a lower-dimensional latent space, significantly reducing the number of features. Notably, our approach preserves the essential characteristics of the original data, ensuring minimal information loss. Lastly, we propose utilizing RNNs/LSTMs as an alternative to Markovian transition models, particularly addressing the limitations associated with the "memoryless" property. By harnessing the sequential nature of RNNs, our framework enables the generation of state transition probabilities with greater user control and flexibility, making it well-suited for real-life applications where memory and context play crucial roles.Overall, our proposed framework offers a comprehensive solution for efficiently analyzing large-scale datasets, empowering researchers and practitioners to extract meaningful insights, make informed decisions, and advance various domains, including finance, healthcare, and engineering.
An Approach to Predict Glacial Lake Outburst Flood
(Indian Statistical Institute, Kolkata, 2022-07) Kayal, Partha
Remote sensing data is a rich resource of information, as it provides a time-wise sequence of data, and therefore can be used for prediction purposes. In this paper, we addressed the challenge of using time series on satellite images to predict the Glacial Lake Outburst Flood(GLOF). In order to predict GLOF, we proposed two-step approach. In the first step, our aim is to extract the pixel-wise information about water, snow, and soil at different time stamps and prepare them for use in the training input. The second step we use is Long Short Term Memory (LSTM) network in order to learn temporal features and thus predict the future pixel value of water, snow, and soil.
View Count Prediction of a Video through Deep Neural Network based Analysis of Subjective Video Attributes
(Indian Statistical Institute, Kolkata, 2020-07) Basu, Spandan
This research work is aimed to design a method for predicting the view count of a video using deep neural network based analysis of subjective video attributes. With more and more companies turning to online video content in uencers to capture the millennial audience, getting people to watch your videos on online platforms is becoming increasingly lucrative. So we provide a solution to the problem by building a model of our own. Our model takes four subjective video attributes as input and predicts the probable view of the video as output. The attributes are the thumbnail image, the title caption, the audio associated with the video and the video itself. We preprocess each of the attributes seperately to obtain the feature vectors. Our model contains four branches to deal with these attributes. We pass the feature vectors of each of the component to the respective branches of the model to capture the salient features with regards to the thumbnail image using a pre-trained CNN architecture, AlexNet; the sentimental feature with regards to the title caption using Sentiment Intensity Analyzer; the temporal feature with regards to the audio waveform using LSTM and both the temporal and salient features with regards to the video using Convolutional LSTM. Since a user, clicks a video based on the title and the thumbnail associated with the video on most online platforms, the model tries to generate a click a nity feature depicting the a nity of the user to click the video. After the user clicked the video, the user decides to view the video based on the audio and the video itself, so the view count of the video is predicted by taking into account the click a nity feature alongwith the temporal feature of the audio waveform and the spatial - temporal feature of the video using a regressor network called the viralvideo- prediction network. A loss function designed from this regression values is used to train the last two stages of the pipeline. We obtain a test accuracy as high as 95.89%.
Exploring neural networks for gesture recognition
(Indian Statistical Institute, Kolkata, 2017) Gupta, Shaunak
Exploring neural networks for gesture recognition
(Indian Statistical Institute, Kolkata, 2017) Gupta, Shaunak

Dissertation and Thesis

Browse

Filters

Settings

Sort By

Results per page

Search Results